Usability and Systems Review for a Swiss Commerce Platform’s AI Module

Context

A Swiss commerce platform’s new AI-visibility module, built to measure how large language models represent a brand against its competitors. Early stage, actively evolving. I joined after the first build was done — the team needed a second pair of eyes with a UX lens, not just a QA check.

My Role

Freelance usability and systems analyst, working async with the lead developer across six review rounds over several weeks. Each round was a live re-test in a real workspace, not a document walkthrough.

Approach

Every finding written in the same format: Observed (what I saw on screen, reproducible) separated from Inferred (what I concluded from it). No merged observations and conclusions in one sentence — that collapse is where most audit reports lose credibility.

Each finding carried a priority and a concrete suggestion in copy the team could paste in. Where a UI issue hinted at something deeper, I cross-checked against the platform’s own backend tools to confirm whether the UI mirrored a real gap in the system logic — not just a missing button.

Selected Findings (generalized)

A status model that collapsed four meanings into one. A list of runs used one red label for everything that wasn’t fully green — partial successes, hard failures, and runs with gaps all read the same. The team split it into four distinct states, color-coded, with counts that now add up. Old runs were retroactively re-labeled.

First-run states that weren’t reachable. Several empty-state flows couldn’t be reproduced because a key setup action was missing from the UI. One action unlocked three findings at once. The resulting empty states now read the actual workspace condition instead of showing generic placeholders.

Degraded-state behavior. With two of three external services returning errors, the product gave three different explanations for the same condition on three different screens. I named the inconsistency, the team unified the copy and the signal.

A missing destructive action. One central object could be created but not removed. Cross-checking the platform’s backend confirmed the missing UI action matched a missing backend verb — this was a systems gap, not a hidden button. The report named it as a scope expansion, not a visual fix.

Delta semantics on a visibility chart. Up/down arrows were applied symmetrically to the user’s own brand and to competitors, which inverts the meaning (a competitor going up is bad for the user). Fix: color deltas only on the user’s own row, with a one-line legend.

Result

Across the first round, 19 of 24 findings were implemented one-to-one, with three solved differently by the team with documented reasoning. The module is live inside the platform today.

Takeaway

For products still finding their shape, the useful question isn’t “does this look right?” It’s “does the screen tell the user what the system actually did?” Most of what moved the module forward wasn’t visual — it was semantic: what a state means, what a label is counting, what an error is actually saying.